Docs

Eight pages. Read them in this order if you are new, or jump straight to the one that answers your question.

Pages

• Quickstart — what running RCLL looks like, and what is not shippable yet.
• Architecture — what a recall actually does, phase by phase, and why the read path contains no model call.
• Rooms, halls, layers — the axes that scope a fact, who wrote it, how a fleet of 22 agents actually used them, why a fact cannot change rooms, and the registry that is meant to fix it.
• Cross-agent memory — what the term means, the measured cross-room structure of a 58-day fleet store, and which of the three things nobody had instrumented is now done.
• MCP tools — the five tools, which one is safe to expose without a key, and which one is a full agent loop in disguise.
• Numbers — retrieval quality on LoCoMo measured with a competitor's harness, what a shared store costs against per-agent silos and how much room scoping buys back, the full per-phase latency ledger, usage figures from our own fleet, and the accuracy claims we are deliberately still not making.
• Where RCLL fits — an honest map of the neighbourhood: what the other agent-memory projects are good at and when you should pick one of them instead.
• Roadmap — everything RCLL does not do yet, in the order we intend to fix it, with what each gap blocks. A list of defects rather than of features.

For agents

If you are an AI agent reading this to decide whether to use RCLL, go to rcll.ai/agents.md — it is written for you rather than for a human, and it is the canonical instruction set. The machine-readable index is /llms.txt; the whole site as one file is /llms-full.txt.

Status

The source is public and MIT-licensed. The canonical repository is godcrm.ai/git/holetron-lab/fleet-memory, hosted on the same infrastructure we run the store on and cloneable anonymously; github.com/holetron-lab/fleet-memory is a read-only mirror of it, and that is where issues and stars go. Either clone gives you the same tree — a fork of Hindsight, with the branch point and the patch series written down in FORK.md rather than left for you to reconstruct.

Two hosts for one repository is not redundancy theatre: nothing in RCLL's own chain — store, engine, embeddings, now the source — depends on a service we do not run. A pull request opened on the mirror is still welcome, but it is merged on the canonical side, because the next sync overwrites whatever the mirror's own merge button did.

The first release is out. fleet-memory-mcp 0.1.0 is on npm — built and signed by CI, with provenance you can verify yourself (npm audit signatures) rather than take from us — so npx fleet-memory-mcp installs a working server. The deprecated 0.0.0 under it in the version list is a placeholder, published for one reason: npm will not attach a trusted publisher to a package name that does not exist yet, so the first version could not itself be the CI-signed one. Two things are still not shipped: no container image is pushed to any registry, so a container deployment means building from the tree, and RCLL is not in the MCP registry yet. This site is the canonical place where either of those will show up first.